Pancreatic cancer is one of the major deadliest cancers, ranking fourth among causes of cancer-related deaths. Pancreatic cancer patients suffer from a poor prognosis with a 5-year survival rate of only 6%. Predicting pancreatic cancer survival is challenging due to different tumor characteristics, treatments and patient populations. Reliable predictions can help in achieving more personalized care and better management. In this study we test the ability of machine learning to predict pancreatic cancer survival.
Pancreatic cancer patients were identified through the Surveillance, Epidemiology and End Results database (2010-2014). Clinical data for patients were extracted including: age, sex, race, tumor site, tumor histology, grade, cancer sequence number, TNM stage, surgery, tumor size, tumor extension, and survival months. Patients’ records were randomly divided into a training set (80%) and a validation set (20%) to predict survival at 6, 12 and 24 months. Different supervised machine learning models were tested to identify models with best predictions.
We identified 14631 patients with median survival of 13 months. Random Forest algorithm achieved better results compared to other tested models. For evaluating model performance, the Area Under the Receiver Operating Characteristic Curve (AUC) of survival prediction was calculated. The trained model yielded AUCs of 85.3% at 6 months, 84.6% at 12 months and 83.2% at 24 months. The most important characteristics which influenced model prediction were: age at diagnosis (19.9%), tumor size (18.5%), surgery (14.6%), and tumor extension (8.4%).Table: 748P
Performance metrics of the trained machined learning model
|Area Under Curve (AUC)||Precision (positive predictive value)||Accuracy||Recall (sensitivity)||F1 Score|
Predicting survival of patients with pancreatic cancer can be achieved using machine learning with good performance of prediction. Improved survival prediction can help in making better treatment decisions and planning social and care needs.
Clinical trial identification
Legal entity responsible for the study
Mohamed H. Osman.
Has not received any funding.
The author has declared no conflicts of interest.